基于CT的影像组学特征同临床物理剂量特征预测肺癌放疗放射性肺炎研究  

Prediction of radiation pneumonia in lung cancer patients by CT-based radiomics signatures and clinical physical dosimetric features

作　　者：陈文韬 孙磊 谭爱斌[3] 唐世强[2] 陈芬[2] 肖建彪 王志芳[2] 甄鑫[1] CHEN Wentao;SUN Lei;TAN Aibin;TANG Shiqiang;CHEN Fen;XIAO Jianbiao;WANG Zhifang;ZHEN Xin(School of Biomedical Engineering,Southern Medical University,Guangzhou 510515,China;Radiotherapy Center,Chenzhou No.1 People's Hospital,Chenzhou 423000,China;Outpatient Department,North Hospital of Chenzhou No.1 People's Hospital,Chenzhou 423000,China)

机构地区：[1]南方医科大学生物医学工程学院,广东广州510515 [2]郴州市第一人民医院放疗中心,湖南郴州423000 [3]郴州市第一人民医院北院门诊,湖南郴州423000

出　　处：《中国医学物理学杂志》2021年第6期672-676,共5页Chinese Journal of Medical Physics

基　　金：国家自然科学基金(81874216);郴州市科技项目(jsyf2017030)。

摘　　要：目的:探讨基于CT的影像组学特征同临床物理剂量特征预测肺癌放疗放射性肺炎研究。方法:回顾性收集2013年1月至2017年1月进行放射治疗的83例肺癌患者的临床物理剂量参数和CT影像以及随访数据。从病例的CT图像中提取107个影像组学特征,结合对应的45个临床物理剂量特征,每例病例共收集152个特征。基于22种特征提取算法和8种分类器构建的176个鉴别模型分析152个特征预测放射性肺炎的准确性以及筛选优势特征的能力。结果:临床物理剂量特征和影像组学特征预测放射性肺炎的鉴别模型中AUC值最高为0.90。前5位的优势特征是:shape_Maximum2DDiameterColumn、shape_Maximum3DDiameter、V_(20)、glcm_Imc1、V_(45)。结论:临床物理剂量特征和影像组学特征通过不同分类器和特征选择算法组合的鉴别模型,可以筛选出理想的鉴别模型以及优势预测特征。Objective To combine CT-based radiomics signatures with clinical physical dosimetric features for predicting radiation pneumonitis in lung cancer patients.Methods The clinical physical dosimetric features,CT images and follow-up data of 83 patients with lung cancer who underwent radiotherapy from January 2013 to January 2017 were retrospectively collected.A total of 152 features,including 107 radiomics signatures extracted from the CT images and 45 clinical physical dosimetric features,were collected for each case.Based on 22 feature extraction methods and 8 classifiers,176 identification models were constructed to analyze the accuracy of 152 features in predicting radiation pneumonia and to evaluate the ability to screen dominant features.Results The highestAUC in the identification model for predicting radiation pneumonitis by clinical physical dosimetric parameter combined with radiomics signatures was 0.90.The top 5 dominant features included shape_Maximum2DDiameterColumn,shape_Maximum3DDiameter,V_(20),glcm_Imc1 and V_(45).Discussion The ideal identification model and superior prediction features can be screened from identification models constructed by the combination of different classifiers and feature selection algorithms based on clinical physical dosimetric features and radiomics signatures.

关 键 词：肺癌 放射性肺炎 特征提取 影像组学 分类器 

分 类 号：R312[医药卫生—基础医学] R818.7